Bank of England rate history example + Windows encoding fix for the examples runner

[TimelordUK]

Recovers and lands work that was in progress when a session crashed. Two independent changes.

1. fix(tests): make the examples runner encoding-safe on Windows

781cdb59 fixed binary resolution on win32, but the suite still depended on PYTHONIOENCODING=utf-8 PYTHONUTF8=1 being set externally. Without them there are three separate cp1252 failures, only the first of which is obvious:

Site	Symptom
printing ✓/✗	UnicodeEncodeError before any results are reported
subprocess.run(text=True)	engine stdout decoded as cp1252 — mangles non-ASCII, breaks formal JSON comparison (formatting_showcase, showcase_deterministic)
open(sql_file, 'r')	UTF-8 example with bytes undefined in cp1252 raises inside get_data_file_hint, the -- #! hint is silently swallowed by the bare except, and the example runs with no data file (solar_system_calculations_simple)

All three are pinned to UTF-8. No behaviour change on Linux/macOS, where UTF-8 was already the locale encoding.

Worth noting the third one fails silently — the hint is dropped and the example proceeds, so it surfaces as a confusing "data file not found" from the engine rather than an encoding error.

2. feat(examples): Bank of England rate history, 1975–2025

258 rate changes from the 1975 post-oil-shock easing through December 2025, tagged with economic eras, the events behind the big moves, and the PM/chancellor/governor of the day.

The raw bankofengland.co.uk download uses a 18 Dec 25 date format the engine treats as a plain string — ORDER BY sorts lexically and YEAR() returns nonsense. scripts/enrich_boe_rates.py rewrites the dates as ISO so a real DATETIME column is inferred, and denormalises the economic context alongside (one data source per invocation, so a join isn't an option).

Both CSVs are committed: the raw download for provenance, the enriched file for the example. Re-running the script reproduces the enriched CSV byte for byte.

The example exercises GROUP BY over derived date parts, CASE banding via CTE, LAG cross-checked against a precomputed column, and date-range filtering. Registered as a formal test with a captured expectation.

Testing

• Examples suite: 32 formal tests pass (up from 31). The 19 smoke failures are pre-existing on main — verified by stashing this work and running the suite against pristine fd04d80 for a baseline. Unchanged by this PR.
• Expectation was originally captured against v1.77.3; re-validated against v1.82.0 after rebuild and it holds unchanged.
• cargo fmt --check clean — no Rust changes in this PR.

🤖 Generated with Claude Code